Driving and braking control method for front and rear axle dual-motor driving pure electric commercial vehicle

Abstract

The invention relates to the technical field of new energy vehicles, in particular to a driving and braking control method for a front and rear axle dual-motor driving pure electric commercial vehicle. Comprising the following steps: acquiring historical working condition data of the dual-motor pure electric vehicle; establishing a vehicle speed prediction model; setting network parameters of the vehicle speed prediction model, and training the prediction model; adjusting network parameters according to the predicted precision to obtain a vehicle speed prediction model; establishing a dual-motor pure electric vehicle model; establishing an improved DDPG proxy model based on a truncated triple Q network and learning rate scheduling; setting a state, an action and a reward function of the improved DDPG proxy model; training a control energy management model based on the training data set; validating the energy management model using the data set; and the trained model is embedded into a whole vehicle controller, and according to the current vehicle speed, the battery electric quantity SOC and the battery health state SOH, the distribution coefficient of the torque of the front axle and the rear axle of the vehicle is calculated in real time to control the operation of the vehicle.

Description

Technical Field

[0001] The present invention relates to the technical field of new energy vehicles, and in particular to a driving and braking control method for a pure electric commercial vehicle driven by dual motors on front and rear axles. Background Art

[0002] As the world pays more and more attention to energy efficiency and environmental protection, new energy vehicles, especially pure electric vehicles, have gradually become the core research direction in the field of transportation. In this context, pure electric commercial vehicles driven by dual motors have attracted widespread attention due to their advantages in improving vehicle dynamic performance, driving stability and energy efficiency management. Compared with the traditional single-motor drive system, the front and rear axle dual-motor drive can achieve more flexible torque distribution, which not only significantly improves the driving experience, but also effectively improves energy utilization. However, the complexity of the dual-motor system also puts higher requirements on the design of efficient drive and braking control strategies.

[0003] Energy management strategy is one of the key technologies to achieve efficient and energy-saving driving and braking control of pure electric vehicles. Traditional energy management methods, such as rule-based control, model predictive control (MPC), and deep reinforcement learning (DRL), have achieved certain results under specific conditions. However, rule-based control strategies show insufficient flexibility in actual driving and are difficult to maintain the optimal operating state of the vehicle under complex and changing working conditions. Although traditional model predictive control methods can predict and optimize future working conditions, they have high computational complexity when facing long prediction cycles and are difficult to respond to the rapid dynamic changes of multi-motor systems in real time. In addition, although a variety of speed prediction algorithms have been studied for vehicle speed prediction, there is still much room for improvement in prediction accuracy. In addition, deep reinforcement learning methods based on artificial intelligence show superiority in dealing with nonlinear and multivariable problems, but due to their lack of systematicity in model predictive control, they are often less adaptable when actual working conditions change.

[0004] Qi Xing and others proposed a torque distribution optimization strategy for front and rear axle dual-motor electric vehicles, and proposed a method to optimize the torque distribution of front and rear axle dual-motor electric vehicles. The article designed two torque distribution strategies: one is a torque distribution method based on motor efficiency optimization, which reasonably distributes the torque of the front and rear axle motors so that the motors work in the optimal efficiency range, thereby improving the working efficiency of the motor system; the second is a torque distribution method based on battery efficiency optimization, which focuses on optimizing battery energy usage and reducing unnecessary battery loss to increase cruising range. However, there is a conflict between these two strategies, because when the motor efficiency is optimal, it may lead to a decrease in battery efficiency, and vice versa. In order to balance the two, the researchers adopted a multi-objective particle swarm optimization algorithm (MOPSO), which can find a suitable torque distribution strategy while considering both motor and battery efficiency;

[0005] First, the computational complexity of the multi-objective particle swarm optimization algorithm (MOPSO) is high. Since the algorithm needs to optimize multiple objectives at the same time, the computational burden of the system may be increased during real-time calculations, resulting in slower processing speeds, which may cause delays in actual driving environments that require rapid responses, affecting the real-time performance of the vehicle. Secondly, this technology only distributes torque based on instantaneous energy consumption, lacks the ability to predict future vehicle operating conditions, and is prone to efficiency decline. Finally, this technology does not adequately monitor the health status of the battery, and is unable to dynamically adjust the energy allocation strategy based on the actual capabilities of the system. Under high power demands or complex driving scenarios, this may lead to the problem of rapid battery health decay. This application uses a prediction model to predict future vehicle speed changes and optimize energy distribution based on future changes in operating conditions. In addition, this application achieves multi-objective optimization of vehicle energy consumption and battery health status by considering the decay of battery life in the reward function of reinforcement learning.

[0006] Wu, Changcheng et al. proposed an energy management strategy for dual-motor pure electric vehicles based on self-attention deep reinforcement learning. In order to cope with the challenges of multiple working conditions and complex road conditions, this technical solution adopts the deep deterministic policy gradient (DDPG) algorithm and combines the gated recurrent unit (GRU) with the self-attention (SA) mechanism to optimize the energy management strategy. Among them, GRU helps the controller to obtain historical state information and improves the training effect of the strategy; the self-attention mechanism enhances the adaptability of the control system to uncertain environments and ensures the optimal energy allocation of the vehicle under variable working conditions. The main optimization goals of this strategy include reducing the aging rate of the battery and reducing the energy consumption of the whole vehicle. By setting multi-objective optimization indicators, the controller can adjust the torque output and energy consumption of the front and rear axle motors in real time to ensure that the overall energy consumption is reduced while extending the battery life.

[0007] However, although certain results have been achieved in the energy management of dual-motor driven pure electric vehicles, this energy management strategy based on deep reinforcement learning is difficult to efficiently cope with complex changes in operating conditions when faced with changeable actual driving conditions. Although this strategy performs well under specific conditions, it cannot dynamically adjust to the optimal energy management strategy when dealing with complex and changing conditions, resulting in a decrease in energy allocation efficiency, which restricts the efficiency improvement and life extension of the system. This application will combine the L-MPC method, under the MPC framework, with a high-precision vehicle speed prediction algorithm, and use the real-time decision-making ability of reinforcement learning to achieve intelligent energy allocation for electric vehicles.

[0008] In order to overcome the above problems, this application proposes a learning-based model predictive control (L-MPC) method, which combines reinforcement learning with model predictive control to achieve complementary advantages of the two methods. Summary of the invention

[0009] The purpose of the present invention is to address the problems existing in the background technology and propose a driving and braking control method for a pure electric commercial vehicle driven by dual motors on the front and rear axles; the purpose of the present invention is to provide an energy management strategy combined with a learning model predictive control (L-MPC) method to achieve intelligent energy distribution for pure electric vehicles driven by dual motors on the front and rear axles, thereby effectively improving the energy utilization efficiency of the system, extending the service life of the battery, and thus improving the overall economy of the vehicle. The present invention aims to solve the deficiencies of the prior art in terms of adaptability, real-time performance and multi-objective optimization, so that pure electric vehicles can still maintain efficient and energy-saving driving and braking control under complex working conditions.

[0010] Specifically, MPC has difficulty handling complex systems and long-term optimization problems, while DRL can cope with the uncertainty of complex systems and long-term scale optimization problems. On the other hand, DRL lacks adaptability to different driving conditions and has poor interpretability, resulting in a significant decrease in control performance when facing new driving conditions. In contrast, the MPC framework has strong interpretability and can obtain local optimal solutions with the help of a speed predictor. Therefore, by combining the two, it is hoped that an efficient energy management strategy can be achieved in complex driving environments. Based on this background, the present application proposes an energy management strategy for pure electric commercial vehicles driven by dual motors on the front and rear axles, which realizes intelligent energy distribution for electric vehicles by combining the L-MPC method. This strategy can not only improve the energy efficiency of the dual-motor system, but also significantly improve the adaptability and operating stability of the vehicle under multiple working conditions, which is of great significance to extending the life of the power system.

[0011] The following is a specific scheme of the present invention:

[0012] A driving and braking control method for a pure electric commercial vehicle driven by dual motors on front and rear axles, comprising the following specific steps:

[0013] S1. Obtain historical operating condition data of dual-motor pure electric vehicles;

[0014] S2. Establish a vehicle speed prediction model based on BiTCN-BiGRU;

[0015] S3, setting network parameters of the vehicle speed prediction model and training the prediction model;

[0016] S4, adjusting the network parameters according to the prediction accuracy and obtaining a high-precision vehicle speed prediction model;

[0017] S5. Establish a dual-motor pure electric vehicle model, including a longitudinal dynamics model, a battery model, and a motor model;

[0018] S6. Establish an improved DDPG agent model based on truncated triple Q network and learning rate scheduling;

[0019] S7. Set the state, action and reward functions of the improved DDPG agent model;

[0020] S8. Based on the training data set, train the energy management model of the learning model predictive control that combines the vehicle speed prediction model with the improved DDPG agent model;

[0021] S9. Use a new operating condition data set to verify the energy management model based on learning model predictive control;

[0022] S10. Embed the model trained in step S9 into the vehicle controller, and calculate the distribution coefficient of the front and rear axle torque of the vehicle in real time according to the current vehicle speed, battery charge SOC, and battery health status SOH to control the vehicle operation.

[0023] Preferably, the historical vehicle operating condition data acquired in step S1 is historical vehicle speed.

[0024] Preferably, in the vehicle speed prediction model trained in step S4, the input data is the current and historical vehicle speed information, and BiTCN performs bidirectional processing on the input data; BiTCN captures the global dependency in the sequence through the bidirectional information flow;

[0025] BiGRU processes the features extracted by BiTCN and combines the forward and backward information flows to model the bidirectional dependencies of the sequence.

[0026] Preferably, the power system of the dual-motor pure electric vehicle model in step S5 includes a first motor, a second motor, a reducer, a clutch, a power battery, a battery management system BMS, a motor control unit, and a DC-DC converter;

[0027] Among them, the power battery is the power source of the entire system, providing electrical energy to drive the vehicle;

[0028] The battery management system BMS is used to monitor the battery status;

[0029] The DC-DC converter is used to convert the high-voltage power of the battery into low-voltage power suitable for other systems in the vehicle;

[0030] The first motor and the second motor drive the tires on the front axle and the rear axle respectively; the two motors are connected to the battery and the motor controller respectively;

[0031] The motor control unit is responsible for controlling the operation of the motor and regulating the motor speed and output torque;

[0032] The reducer is used to reduce the output speed of the motor and increase the output torque to meet the rotation requirements of the wheels;

[0033] The clutch is used to disconnect the transmission connection between the motor and the tire.

[0034] Preferably, the driving torque required by the vehicle in step S5 is calculated by longitudinal dynamics:

[0035]

[0036] in is the conversion factor of the rotating mass; is the vehicle mass; is the driving acceleration; is the slope angle, and the rolling resistance coefficient is , the acceleration due to gravity is , the air resistance coefficient is Indicates; wheel radius is , the speed is , the frontal area is ; The influence of slope is not considered, so ; Then the above formula becomes:

[0037]

[0038] The two motors meet the required torque of the vehicle through torque coupling, based on the torque distribution factor , the torque of the front and rear motors is calculated as follows:

[0039]

[0040] In the formula, represents the output torque of the front motor, represents the output torque of the rear motor; in addition, when , only the rear motor provides power, when , only the front motor provides power; when , realize front and rear drive, and the speed of the two motors is consistent at this time;

[0041] Model the motor; the efficiency of the motor in the driving state and the braking recovery state is different, and the output power of the motor is calculated as follows:

[0042]

[0043] also, Indicates the efficiency of the motor, which is queried by the efficiency map, that is, ;

[0044] Modeling of power batteries, including equivalent internal resistance model and battery aging model;

[0045] The battery SOC and power in the equivalent internal resistance model are derived from the following formula:

[0046]

[0047] in, is the output power of the battery, is the circuit current, is the open circuit voltage, is the internal resistance of the battery, is the battery capacity, is the initial SOC;

[0048] Secondly, build a battery aging model; the battery health (SOH) formula is expressed as follows:

[0049]

[0050] in, is the equivalent number of cycles before the battery pack reaches the end of its life, is the discharge rate. The above formula is modified into a discrete form and the formula is expressed as follows:

[0051]

[0052] in, is the current duration; the capacity loss empirical model based on the Arrhenius equation calculates the influence of discharge rate and internal temperature, and the formula is expressed as follows:

[0053]

[0054] in, is the percentage of capacity loss, represents the pre-exponential factor, is the ideal gas constant equal to 8.314 J / (mol•K), is a power law factor equal to 0.55, Indicates the ampere-hour throughput; It represents the activation energy in J / mol, and the formula is as follows:

[0055]

[0056] when When it drops to 20%, the vehicle power battery pack reaches the end of its life. At this time, the formula for the ampere-hour throughput and the equivalent number of cycles before the power battery pack reaches the end of its life is expressed as follows:

[0057]

[0058]

[0059] Combine the given current, temperature and battery dynamics to calculate the battery health; in the above formula is the average temperature inside the battery.

[0060] At the same time, the powertrain components satisfy the power balance equation:

[0061]

[0062] in, is the output power of the battery, represents the output power of the first motor, Represents the output power of the second motor.

[0063] Preferably, the DDPG network structure in step S6 consists of four networks: the current Actor network, the target Actor network, the current Critic network and the target Critic network;

[0064] The current Actor network is used to generate strategies , the current Critic network generates a Q value to evaluate the current strategy; the experience replay pool stores the data samples generated by each interaction between the agent and the environment , and learn by random mini-batch sampling;

[0065] The current Actor network is based on the current state , select the current moment action , used to interact with the environment to generate the state of the next moment and reward value ; The target Actor network is responsible for the next state Select next action ;

[0066] Since the weights of the target network are periodically updated from the current network, the target Q value can be temporarily fixed during training, making learning more stable; the target value of the Q value:

[0067]

[0068] in, is the reward value at the current time step t, is the discount factor, which indicates the impact of future rewards on long-term benefits, and its value range is [0, 1]; Generated by the target Critic network value; The strategy function for the target Actor network is its network weight; is the weight of the target Critic network;

[0069] The current goal of the Critic network is to minimize the loss function. The loss function used in network training is:

[0070]

[0071] Where K: the number of samples in a batch (i.e., mini-batch size), is the target Q value calculated by the target Critic network; The Q value generated by the current Critic network; is the weight of the current Critic network;

[0072] The current goal of the Actor network is to maximize the expected return, that is

[0073]

[0074] In the formula, is the strategy function of the current Actor network;

[0075] The weight gradient of the current Actor network is:

[0076]

[0077] is the gradient of the Critic network’s evaluation of the action value, which measures How does a small change in The value of Actor Network Strategy The gradient of Adjusting parameters To optimize the selected action; through the chain rule, the gradient of the Actor network is ultimately determined by the product of the value gradient of the Critic network and the Actor policy gradient, guiding the Actor network to update the weight so that the selected action can better improve the cumulative benefit;

[0078] Use the optimizer to update the weights of the current Actor network and the weight of the current evaluation network , and then use the soft update method to update the weights of the two target networks respectively, namely:

[0079]

[0080] In the formula, It is a soft update parameter.

[0081] Preferably, in step S6, a truncated triple Q learning strategy is adopted to obtain a more accurate Q value estimation;

[0082] The objective function expression is as follows:

[0083]

[0084] In the formula, and Represents different network parameters and The Q output under is the next state,

[0085] is a set of possible actions; mean represents the mean operation on the truncated Q value.

[0086] 8. The driving and braking control method for a pure electric commercial vehicle driven by dual motors on front and rear axles according to claim 1 is characterized in that, in step S7, a CyclicLR learning rate scheduler is introduced to adjust the learning rate for periodic decay;

[0087] The reward function is defined as follows:

[0088]

[0089] And, define the boundaries in the learning process as follows:

[0090]

[0091] In the formula, and Indicates the upper and lower limits of the battery SOC, and Indicates the upper and lower limits of the battery health status. Indicates the upper limit of the first motor speed, Indicates the upper limit of the second motor speed, Indicates the upper limit of the torque of the first motor, Indicates the upper limit of the torque of the second motor.

[0092] Preferably, in step S8, assuming that the vehicle speed prediction time domain is 3s, at the current time k, the vehicle speed is v, the historical vehicle speeds at k-3, k-2, k-1 and the current vehicle speed at time k are input into the vehicle speed prediction module, and the module outputs the predicted vehicle speeds at time k+1, k+2, k+3; then The vehicle speed information is input into the intelligent agent module of the learning predictive control for training, and the vehicle speed is used , battery SOC, and SOH update the dual-motor power system status within the prediction time domain:

[0093]

[0094] Get the optimal action at time t Then, using the state quantity at time k and the best action , get the reward at time k , and the next state quantity , then the sequence Add to the sequence of experiences;

[0095] The strategy network and value network of the improved DDPG algorithm are used for training to obtain the energy management model of dual-motor pure electric vehicles based on learning model predictive control.

[0096] Preferably, in step S10, after the training is completed, the energy management model of the learning predictive control is loaded into the vehicle controller, and the specific execution is as follows:

[0097] During the actual driving process of the vehicle, the short-term historical speed of the vehicle and the current speed of the vehicle are input into the speed prediction model to solve the speed sequence of the vehicle in the prediction time domain in real time;

[0098] Then, the vehicle speed in the predicted time domain is input into the trained DDPG agent. At each moment, the optimal control action is obtained through calculation, including the torque distribution coefficient of the front and rear motors, and it is input into the motor controller to distribute the front and rear motor torques, realizing real-time rolling optimization.

[0099] Compared with the prior art, the present invention has the following beneficial technical effects:

[0100] The technical solution of the present invention realizes intelligent and economical energy distribution and health status perception of pure electric vehicles by introducing a dual-motor energy management strategy based on learning model predictive control (L-MPC). The specific beneficial effects brought by the solution include:

[0101] 1. Improve adaptability to working conditions

[0102] The present invention predicts the vehicle speed by using a vehicle speed prediction model, so that the energy management strategy can be adaptively adjusted according to future driving conditions. This solution shows extremely high adaptability when facing complex and dynamically changing road conditions and driving needs, ensuring that the system can adjust the power distribution of the front and rear axle motors according to different working conditions, thereby achieving the best energy consumption performance and driving stability.

[0103] 2. Realize real-time response and improve control stability

[0104] This invention combines L-MPC and deep reinforcement learning (DDPG) to overcome the computational complexity of traditional model predictive control under complex working conditions and achieve efficient real-time response within a long prediction cycle. The strategy update provided by the DDPG algorithm enables the system to quickly respond to changes in driving conditions, while improving vehicle economy, ensuring the real-time and stability of energy management, and bringing a smoother driving experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0105] Figure 1 This is a diagram of the energy management system of a dual-motor pure electric vehicle based on learning model predictive control;

[0106] Figure 2 This is the structural diagram of the chassis of a pure electric dual-motor vehicle;

[0107] Figure 3 Schematic diagram of learning rate scheduling;

[0108] Figure 4 This is a bidirectional causal expansion convolution structure diagram;

[0109] Figure 5 This is the BiTCN algorithm structure diagram;

[0110] Figure 6 It is the GRU structure diagram;

[0111] Figure 7 This is a short-term vehicle speed prediction framework diagram;

[0112] Figure 8 A framework for training energy management strategies for dual-motor pure electric vehicles with learning-based predictive control. DETAILED DESCRIPTION

[0113] Example 1

[0114] The present invention provides a dual-motor energy management strategy based on learning-based model predictive control (L-MPC) for pure electric vehicles with dual motors on the front and rear axles. This solution combines deep reinforcement learning algorithms, multi-objective optimization, and high-precision vehicle speed prediction methods to achieve intelligent energy distribution and health status perception of electric vehicles, thereby improving the energy utilization efficiency of the system and extending the system life. The specific technical solution is as follows Figure 1 shown.

[0115] The technical solution of this embodiment proposes a dual-motor pure electric vehicle energy management method based on learning model predictive control, which is optimized by combining vehicle speed prediction with reinforcement learning strategy. The technical solution includes a training phase and an application phase. In the training phase, a longitudinal dynamics model and a battery model are established, and an improved deep deterministic policy gradient (DDPG) agent model based on a truncated triple Q network and learning rate scheduling is set, and the state, action and reward function of the DDPG agent model are set. At the same time, the historical operating data of the dual-motor pure electric vehicle is obtained, and a high-precision vehicle speed prediction model is trained using a bidirectional convolutional time network-bidirectional gated recurrent neural network (BiTCN-BiGRU) architecture. On this basis, the vehicle speed prediction is combined with reinforcement learning to construct a learning-based predictive control energy management model.

[0116] During the application stage, the trained model is embedded in the vehicle controller, and the distribution coefficient of the vehicle's front and rear axle torque is calculated in real time based on observable variables such as current vehicle speed, battery charge SOC, and battery health status SOH, so as to achieve efficient energy distribution and adaptability to driving conditions.

[0117] In the training phase, the specific modeling process is as follows:

[0118] 1. Power system modeling

[0119] The model built in this application is a pure electric dual-motor power system. Figure 2 , the power system includes the first motor, the second motor, the reducer, the clutch, the power battery, the battery management system (BMS), the motor control unit, the DC-DC converter and other main components. Among them, the power battery is the power source of the entire system, providing electric energy to drive the vehicle. The battery management system is used to monitor the battery status to ensure that it works under safe and efficient conditions. The DC-DC converter is used to convert the high-voltage electric energy of the battery into low-voltage electric energy suitable for other systems of the vehicle. The first motor and the second motor drive the tires on the front axle and the rear axle respectively. The two motors are connected to the battery and the motor controller respectively. The motor control unit is responsible for controlling the operation of the motor and adjusting the speed and output torque of the motor. The reducer is used to reduce the output speed of the motor and increase the output torque to meet the rotation requirements of the wheel. The clutch is used to disconnect the transmission connection between the motor and the tire. During the operation of the vehicle, the power battery provides energy to the motor, converts the high speed of the motor into a suitable torque through the reducer, and then transmits the torque to the drive wheel through the clutch, so that the vehicle moves forward smoothly. At the same time, the motor control unit dynamically adjusts the torque distribution of the front and rear axle motors to adapt to different road conditions and driving requirements, improve energy efficiency and enhance maneuverability, and the regenerative braking system recovers energy during deceleration to further extend the driving range. This front and rear axle dual-motor system can achieve flexible power distribution, and the torque distribution between the front and rear axles can be adjusted according to road conditions, driving modes and other requirements to improve energy efficiency.

[0120] Next, the system and components are modeled, and the required driving torque of the vehicle is calculated from the longitudinal dynamics:

[0121]

[0122] in is the conversion factor of the rotating mass; is the vehicle mass; is the driving acceleration; is the slope angle, and the rolling resistance coefficient is , the acceleration due to gravity is , the air resistance coefficient is Indicates; wheel radius is , the speed is , the frontal area is ; The influence of slope is not considered, so ; Then the above formula becomes:

[0123]

[0124] In addition, for the dual-motor system of the present application, the two motors meet the required torque of the vehicle through torque coupling, based on the torque distribution factor , the torque of the front and rear motors is calculated as follows:

[0125]

[0126] In the formula, represents the output torque of the front motor, represents the output torque of the rear motor; in addition, when , only the rear motor provides power, when , only the front motor provides power; when , realizing front and rear drive, and at this time the speeds of the two motors are consistent.

[0127] Next, we will model the motor. The motor has different efficiencies in the driving state and the braking recovery state. The motor output power is calculated as follows:

[0128]

[0129] also, Indicates the efficiency of the motor, which is queried by the efficiency map, that is, .

[0130] Next, we will model the power battery. The charging and discharging process of the power battery is a very complex electrochemical reaction process, and it is very important to establish an accurate battery model. The power battery model built in this application contains two sub-models, the equivalent internal resistance model and the battery aging model.

[0131] First, the battery SOC and power in the equivalent internal resistance model are derived from the following formula:

[0132]

[0133] in, is the output power of the battery, is the circuit current, is the open circuit voltage, is the internal resistance of the battery, is the battery capacity, is the initial SOC.

[0134] Secondly, build a battery aging model. First, the battery health formula is expressed as follows:

[0135]

[0136] in, is the equivalent number of cycles before the battery pack reaches the end of its life. The above formula is modified into a discrete form and expressed as follows:

[0137]

[0138] in, is the duration of the current. The capacity loss empirical model based on the Arrhenius equation calculates the influence of discharge rate and internal temperature, and the formula is as follows:

[0139]

[0140] in, is the percentage of capacity loss, represents the pre-exponential factor, is the ideal gas constant equal to 8.314 J / (mol•K), is a power law factor equal to 0.55, Indicates the ampere-hour throughput; It represents the activation energy in J / mol, and the formula is as follows:

[0141]

[0142] when When it drops to 20%, the vehicle power battery pack reaches the end of its life. At this time, the formula for the ampere-hour throughput and the equivalent number of cycles before the power battery pack reaches the end of its life is expressed as follows:

[0143]

[0144] in, is the duration of the current. The battery health is calculated by combining the given current, temperature and battery dynamics using formula (5).

[0145] At the same time, the powertrain components satisfy the power balance equation:

[0146]

[0147] in, is the output power of the battery, represents the output power of the first motor, Represents the output power of the second motor.

[0148] 2. Reinforcement Learning Algorithms

[0149] 2.1 DDPG Algorithm

[0150] The DDPG algorithm is a reinforcement learning algorithm for continuous action space, using Actor-Critic as its basic framework. The DDPG network structure consists of four networks: the current Actor network, the target Actor network, the current Critic network, and the target Critic network.

[0151] The current Actor network is used to generate strategies , the current Critic network generates a Q value to evaluate the current strategy. The experience replay pool stores the data samples generated by each interaction between the agent and the environment. , and learns through random small batch sampling. The current Actor network is based on the current state , select the current moment action , used to interact with the environment to generate the state of the next moment and reward value ; The target Actor network is responsible for the next state Select next action Since the weights of the target network are periodically updated from the current network, the target Q value can be temporarily fixed during training, making learning more stable. Target value of Q value:

[0152]

[0153] in, is the reward value at the current time step t, is the discount factor, which indicates the impact of future rewards on long-term benefits, and its value range is [0, 1]; Generated by the target Critic network value; The strategy function for the target Actor network is its network weight; is the weight of the target Critic network.

[0154] The current goal of the Critic network is to minimize the loss function. The loss function used in network training is:

[0155]

[0156] Where K: the number of samples in a batch (i.e., mini-batch size), is the target Q value calculated by the target Critic network; The Q value generated by the current Critic network; is the weight of the current Critic network.

[0157] The current goal of the Actor network is to maximize the expected return, that is

[0158]

[0159] In the formula, Is the strategy function of the current Actor network.

[0160] The weight gradient of the current Actor network is:

[0161]

[0162] is the gradient of the Critic network’s evaluation of the action value, which measures How does a small change in The value of Actor Network Strategy The gradient of Adjusting parameters To optimize the selected action. Through the chain rule, the gradient of the Actor network is ultimately determined by the product of the value gradient of the Critic network and the Actor policy gradient, guiding the Actor network to update the weights so that the selected action can better improve the cumulative benefit.

[0163] Use the optimizer to update the weights of the current Actor network and the weight of the current evaluation network , and then use the soft update method to update the weights of the two target networks respectively, namely:

[0164]

[0165] In the formula, It is a soft update parameter.

[0166] 2.2 Truncated Triple Q-Learning Mechanism

[0167] In the DDPG algorithm, the target critic network uses the maximum value as the expected value to obtain its Q value, which is generally overestimated relative to the true expected Q value. The overestimation of the value function in DDPG generally leads to two problems: On the one hand, the overestimation of the value will lead to significant deviations after multiple updates. On the other hand, the deviation in value estimation will further lead to inaccurate policy updates. The critic network overestimates the suboptimal action, which will lead the policy network to choose the suboptimal action during the policy update.

[0168] In order to alleviate the problem of overestimation of Q value in DDPG algorithm, some studies have designed a dual evaluation network mechanism, the core of which is to reduce the risk of overestimation of Q value by correcting each other through two independent Q networks. However, this method can reduce the risk of overestimation, but sometimes underestimates. Therefore, this application adopts a truncated triple Q learning strategy, the main idea of ​​which is to combine the bias of overestimation and underestimation at the same time to obtain a more accurate Q value estimate.

[0169] The objective function expression of this method is as follows:

[0170]

[0171] In the formula, and Represents different network parameters and The Q output under is the next state,

[0172] is a set of possible actions; mean represents the mean operation on the truncated Q value.

[0173] This mechanism considers the minimum value of multiple Q network outputs at the same time and takes the average to further smooth the result, thereby reducing the risk of overestimation while avoiding the negative impact of underestimation on policy optimization. This truncated triple Q learning strategy further improves the accuracy of Q value estimation while retaining the advantages of the dual evaluation network, thereby more effectively guiding policy network optimization.

[0174] 2.3 Learning Rate Scheduling

[0175] The traditional learning rate (i.e. fixed learning rate) has shown certain limitations when training deep reinforcement learning models. The main feature of the fixed learning rate is that the step size remains unchanged during the training process. This method has the advantages of being intuitive and easy to implement, but it also has problems such as limited convergence speed, local optimality, and overfitting in complex learning tasks. The specific manifestations are as follows:

[0176] First, a fixed learning rate makes it difficult to achieve fast convergence in the early stages of training. If the learning rate is too low, the model will take longer to explore the parameter space and reach the optimal solution; but if the learning rate is too high, the model may skip the optimal solution when updating, resulting in unstable training.

[0177] Secondly, in deep reinforcement learning, especially in policy network optimization, a fixed learning rate can easily cause the model to be trapped in a local optimal solution and make it difficult to escape.

[0178] Finally, the fixed learning rate is not decayed in the late stages of training, causing the model to overfit the training data and not generalize well to new data.

[0179] In order to solve these problems, this application introduces a learning rate scheduler in the reinforcement learning algorithm. The learning rate scheduler can dynamically adjust the learning rate during the training process, making the learning process more stable and effective, and easier to find the global optimal solution. Among them, CyclicLR is a scheduler that periodically adjusts the learning rate, allowing the learning rate to fluctuate between the minimum and maximum values, helping the model to quickly explore in the early stages of training and gradually converge in the later stages.

[0180] Specifically, this application adopts a learning rate scheduler with periodic fluctuations and peak decay, as follows Figure 3 As shown in Figure 1. In each cycle, the learning rate gradually increases from the lowest value to the highest value, and then drops back to the lowest value, forming a triangular fluctuation pattern. At the beginning of each new cycle, the highest learning rate is halved compared to the previous cycle. The process is as follows:

[0181] The first cycle: The learning rate gradually increases from the lowest value to the maximum value, and then falls back to the lowest value, forming a complete triangular waveform.

[0182] Second cycle: The learning rate increases from the lowest value to half of the maximum value of the first cycle, and then decreases to the lowest value.

[0183] The third cycle: The maximum learning rate is halved again, the learning rate rises from the lowest value to this new highest value, and then falls back to the lowest value.

[0184] As the training cycle progresses, the fluctuation of the learning rate gradually decreases and eventually stabilizes in a lower range. This mode can effectively conduct greater exploration in the early stages of training, and gradually reduce the fluctuation of the learning rate in the later stages of training, allowing the model to converge more stably.

[0185] 3. BiTCN-BiGRU vehicle speed prediction algorithm

[0186] 3.1 Bidirectional Convolutional Temporal Network

[0187] Bidirectional Convolutional Temporal Network (BiTCN) is a deep learning model designed for processing sequence data. BiTCN effectively improves the prediction effect of sequence data by combining the characteristics of temporal convolutional networks and the bidirectional information flow mechanism. Among them, the temporal convolutional network uses causal convolution to ensure that each element of the output sequence depends only on the current and previous inputs, avoiding the influence of future information on current decisions. This causal constraint ensures the rationality of the prediction and simulates the actual characteristics of the time series to a certain extent. In addition, BiTCN uses dilated convolution to enable the model to capture longer historical information, thereby improving performance when dealing with long-term dependencies.

[0188] like Figure 4 This is the structural diagram of the bidirectional causal expansion convolution. By introducing a bidirectional information flow mechanism, the model can simultaneously consider historical information. Input and future information This bidirectional design allows the model to learn not only from current and past data, but also to integrate future information, thereby modeling sequence data more comprehensively. In addition, as shown in the figure, the receptive field of the dilated convolution can be significantly increased by sampling at intervals, thereby obtaining a larger receptive field using fewer network layers.

[0189] The output calculation formula of BiTCN, which combines bidirectional information flow and causal convolution at time t, is:

[0190]

[0191] in, is the weight of the convolution kernel, represents the input item after interval sampling, is the size of the convolution kernel, and d is the dilation coefficient.

[0192] The BiTCN algorithm is mainly composed of a causal convolution layer, an activation function, a weight normalization layer, and a spatial Dropout layer to achieve prediction of time series data.

[0193] 3.2 Bidirectional Gated Recurrent Neural Network

[0194] Gated Recurrent Unit (GRU) GRU unit consists of a reset gate and an update gate to control the flow of information. The reset gate determines how much past information is forgotten in the current state. The update gate determines how much information in the current state needs to be retained and passed to the next time step. The structure of the GRU unit is as follows Figure 6 shown.

[0195] The calculation formula of its status and output is as follows:

[0196]

[0197] In the formula, is the input vector at the current moment, is the hidden state at the previous moment, is the hidden state at the current moment, It is the update gate. It is the reset gate. is the new information generated at time t. represents the input weight matrix,

[0198] A weight matrix representing the hidden state.

[0199] The Bidirectional Gated Recurrent Unit (BiGRU) combines the characteristics of bidirectional recurrent neural networks and gated recurrent units. It can simultaneously consider the forward and backward information flows of data during processing, thereby enhancing the ability to capture global dependencies and performing well in time series learning tasks.

[0200] 3.3 Short-term vehicle speed prediction framework

[0201] This application combines the advantages of BiTCN and BiGRU to achieve accurate prediction of future vehicle speed sequences. Among them, the input data is the current and historical vehicle speed information. BiTCN processes the input data bidirectionally, and uses its causal convolution and dilated convolution characteristics to expand the receptive field, ensuring that the model does not lose the time order when capturing historical information. Through bidirectional information flow, BiTCN can capture the global dependencies in the sequence without increasing the computational complexity, helping the model understand the overall trend and change pattern of the sequence. Subsequently, BiGRU further processes the features extracted by BiTCN, combines the forward and backward information flows, and realizes the modeling of bidirectional dependencies of the sequence. Because BiGRU has fewer parameters and higher computational efficiency, it has advantages in vehicle speed prediction tasks with higher real-time requirements. The bidirectional structure allows the model to better understand the trend of vehicle speed changes, and is particularly suitable for capturing complex time dependencies.

[0202] Energy management strategy for dual-motor pure electric vehicles based on learning predictive control

[0203] 4.1 Reinforcement Learning Setup

[0204] In the deep reinforcement learning algorithm, the agent observes the state of the environment and takes corresponding actions to obtain immediate rewards and transfer to a new state. The design of the state space, action space and reward function in this process is very important. In the energy management of dual-motor pure electric vehicles, the agent needs to obtain the vehicle's motion state and the battery state. Specifically, the vehicle's motion state includes the vehicle speed (v), acceleration (a) and the torque required for the vehicle. The battery state includes the battery state of charge (SOC) and health state (SOH), which are used to characterize the current power and health of the battery. Therefore, the definition of the state space includes: .

[0205] The action space of reinforcement learning defines the actions taken by the agent at each time step. In this application, the torque distribution factor Defined as an action space for controlling the torque distribution between the first motor and the second motor.

[0206] In addition, the core idea of ​​reinforcement learning is that through interaction with the environment, the agent learns a strategy that can obtain the maximum expected reward. Therefore, designing a reasonable reward function is very important for the training of the agent. In this application, the optimization objectives include two aspects: the first is to reduce the power consumption of the battery, by minimizing the power consumption of the power battery to improve energy efficiency; the second is to reduce the health decline of the battery: by optimizing energy distribution, the battery life is extended. Therefore, the reward function is defined as follows:

[0207]

[0208] And, define the boundaries in the learning process as follows:

[0209]

[0210] 4.2 Energy Management Training Process of Learning Predictive Control

[0211] For the control strategy proposed in this application, the specific training framework is as follows Figure 8 As shown:

[0212] Assuming that the speed prediction time domain is 3s, at the current time k, the vehicle speed is v, the historical speeds of k-3, k-2, k-1 and the current speed at time k are input into the speed prediction module, which outputs the predicted speeds at time k+1, k+2, and k+3. The vehicle speed information is input into the reinforcement learning agent module and goes through the following steps:

[0213] Table 1. Algorithm execution process

[0214]

[0215] Using the vehicle's speed , battery SOC, SOH, calculate and predict the state of the dual-motor power system in the time domain:

[0216]

[0217] Get the optimal action at time t Then, using the state quantity at time k and the best action , get the reward at time k , and the next state quantity , then the sequence Add to the experience sequence.

[0218] Then, the above (10)~(14) are used to train the strategy network and value network of the improved DDPG algorithm to obtain the energy management strategy of the dual-motor pure electric vehicle based on learning model predictive control.

[0219] 4.3 Energy management execution process of learning-based predictive control

[0220] After the training is completed, the energy management strategy of the learning predictive control is loaded into the vehicle controller. The specific execution is as follows:

[0221] During the actual driving process of the vehicle, the short-term historical speed of the vehicle and the current speed of the vehicle are input into the speed prediction model to solve the speed sequence of the vehicle in the prediction time domain in real time;

[0222] Then, the vehicle speed in the predicted time domain is input into the trained DDPG agent. At each moment, the optimal control action (torque distribution coefficient of the front and rear motors) is obtained through the algorithm process in Table 1 and input into the motor controller to distribute the front and rear motor torques, thereby achieving real-time rolling optimization.

[0223] In this embodiment, 1. The traditional DDPG reinforcement learning algorithm is susceptible to the influence of overestimation bias, and the general improvement method is to correct each other through two independent Q networks to reduce the risk of overestimation of Q value. However, this improvement method can reduce the risk of overestimation, but sometimes underestimation occurs. The present application adopts a strategy of truncated triple Q learning. The main idea is to combine the overestimation and underestimation biases at the same time, introduce three Q networks to estimate the action value, and obtain a more accurate Q value estimate through a truncation mechanism. In addition, the learning rate scheduling proposed in the present application enables greater exploration in the early stage of training, and gradually reduces the fluctuation of the learning rate in the later stage of training, so that the model can converge more stably and improve the performance of the algorithm.

[0224] 2. The vehicle speed prediction model based on BiTCN and BiGRU combines the advantages of convolutional neural networks and recurrent neural networks, and can efficiently capture long-term dependencies and local features in time series data, thereby improving the accuracy and robustness of vehicle speed prediction.

[0225] 3. The present invention utilizes the framework of model predictive control and introduces a vehicle speed prediction algorithm to accurately predict future driving conditions. Combined with the decision-making ability of reinforcement learning, the system can adaptively adjust the torque output of the front and rear axle dual motors according to different operating conditions, thereby ensuring that the vehicle maintains efficient energy distribution in a complex driving environment and achieves better energy consumption performance.

[0226] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto, and various changes can be made within the knowledge scope of technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. A driving and braking control method for a pure electric commercial vehicle driven by dual motors on front and rear axles, characterized in that: The specific steps include: S1. Obtain historical operating condition data of dual-motor pure electric vehicles; S2. Establish a vehicle speed prediction model based on BiTCN-BiGRU; S3, setting network parameters of the vehicle speed prediction model and training the prediction model; S4, adjusting the network parameters according to the prediction accuracy and obtaining a high-precision vehicle speed prediction model; S5. Establish a dual-motor pure electric vehicle model, including a longitudinal dynamics model, a battery model, and a motor model; S6. Establish an improved DDPG agent model based on truncated triple Q network and learning rate scheduling; S7. Set the state, action and reward functions of the improved DDPG agent model; S8. Based on the training data set, train the energy management model of the learning model predictive control that combines the vehicle speed prediction model with the improved DDPG agent model; S9. Use a new operating condition data set to verify the energy management model based on learning model predictive control; S10. Embed the model trained in step S9 into the vehicle controller, and calculate the distribution coefficient of the front and rear axle torque of the vehicle in real time according to the current vehicle speed, battery charge SOC, and battery health status SOH to control the vehicle operation.

2. The driving and braking control method for a pure electric commercial vehicle driven by dual motors on front and rear axles according to claim 1, characterized in that: The historical vehicle operating condition data acquired in step S1 is the historical vehicle speed.

3. The driving and braking control method for a pure electric commercial vehicle driven by dual motors on front and rear axles according to claim 1, characterized in that: In the vehicle speed prediction model trained in step S4, the input data is the current and historical vehicle speed information, and BiTCN processes the input data bidirectionally; BiTCN captures the global dependency in the sequence through bidirectional information flow; BiGRU processes the features extracted by BiTCN and combines the forward and backward information flows to model the bidirectional dependencies of the sequence.

4. The driving and braking control method for a pure electric commercial vehicle driven by dual motors on front and rear axles according to claim 1, characterized in that: In step S5, the power system of the dual-motor pure electric vehicle model includes a first motor, a second motor, a reducer, a clutch, a power battery, a battery management system BMS, a motor control unit, and a DC-DC converter; Among them, the power battery is the power source of the entire system, providing electrical energy to drive the vehicle; The battery management system BMS is used to monitor the battery status; The DC-DC converter is used to convert the high-voltage power of the battery into low-voltage power suitable for other systems in the vehicle; The first motor and the second motor drive the tires on the front axle and the rear axle respectively; the two motors are connected to the battery and the motor controller respectively; The motor control unit is responsible for controlling the operation of the motor and regulating the motor speed and output torque; The reducer is used to reduce the output speed of the motor and increase the output torque to meet the rotation requirements of the wheels; The clutch is used to disconnect the transmission connection between the motor and the tire.

5. The driving and braking control method for a pure electric commercial vehicle driven by dual motors on front and rear axles according to claim 4 is characterized in that: In step S5, the driving torque required by the vehicle is calculated by the longitudinal dynamics: in is the conversion factor of the rotating mass; is the vehicle mass; is the driving acceleration; is the slope angle, and the rolling resistance coefficient is , the acceleration due to gravity is , the air resistance coefficient is Indicates; wheel radius is , the speed is , the frontal area is ; The influence of slope is not considered, so ; then the above formula becomes: The two motors meet the required torque of the vehicle through torque coupling, based on the torque distribution factor , the torque of the front and rear motors is calculated as follows: In the formula, represents the output torque of the front motor, represents the output torque of the rear motor; in addition, when , only the rear motor provides power, when , only the front motor provides power; when , realize front and rear drive, and the speed of the two motors is consistent at this time; Model the motor; the efficiency of the motor in the driving state and the braking recovery state is different, and the output power of the motor is calculated as follows: also, Indicates the efficiency of the motor, which is queried by the efficiency map, that is, ; Modeling of power batteries, including equivalent internal resistance model and battery aging model; The battery SOC and power in the equivalent internal resistance model are derived from the following formula: in, is the output power of the battery, is the circuit current, is the open circuit voltage, is the internal resistance of the battery, is the battery capacity, is the initial SOC; Secondly, build a battery aging model; the battery health SOH formula is expressed as follows: in, is the equivalent number of cycles before the battery pack reaches the end of its life, is the discharge rate. The above formula is modified into a discrete form and the formula is expressed as follows: in, is the duration of the current; the capacity loss empirical model based on the Arrhenius equation calculates the influence of discharge rate and internal temperature, and the formula is expressed as follows: in, is the percentage of capacity loss, represents the pre-exponential factor, is the ideal gas constant equal to 8.314 J / (mol•K), is a power law factor equal to 0.55, Indicates the ampere-hour throughput; It represents the activation energy in J / mol, and the formula is as follows: when When it drops to 20%, the vehicle power battery pack reaches the end of its life. At this time, the formula for the ampere-hour throughput and the equivalent number of cycles before the power battery pack reaches the end of its life is expressed as follows: Combine the given current, temperature and battery dynamics to calculate the battery health; in the above formula is the average temperature inside the battery. At the same time, the powertrain components satisfy the power balance equation: in, is the output power of the battery, represents the output power of the first motor, Represents the output power of the second motor.

6. The driving and braking control method for a pure electric commercial vehicle driven by dual motors on front and rear axles according to claim 1, characterized in that: In step S6, the DDPG network structure consists of four networks: the current Actor network, the target Actor network, the current Critic network, and the target Critic network; The current Actor network is used to generate strategies , the current Critic network generates a Q value to evaluate the current strategy; the experience replay pool stores the data samples generated by each interaction between the agent and the environment , and learn by random mini-batch sampling; The current Actor network is based on the current state , select the current moment action , used to interact with the environment to generate the state of the next moment and reward value ; The target Actor network is responsible for the next state Select next action ; Since the weights of the target network are periodically updated from the current network, the target Q value can be temporarily fixed during training, making learning more stable; the target value of the Q value: in, is the reward value at the current time step t, is the discount factor, which indicates the impact of future rewards on long-term benefits, and its value range is [0, 1]; Generated by the target Critic network value; The strategy function for the target Actor network is its network weight; is the weight of the target Critic network; The current goal of the Critic network is to minimize the loss function. The loss function used in network training is: Where K: the number of samples in a batch, is the target Q value calculated by the target Critic network; The Q value generated by the current Critic network; is the weight of the current Critic network; The current goal of the Actor network is to maximize the expected return, that is In the formula, is the strategy function of the current Actor network; The weight gradient of the current Actor network is: is the gradient of the Critic network’s evaluation of the action value, which measures How does a small change in The value of Actor Network Strategy The gradient of Adjusting parameters To optimize the selected action; Use the optimizer to update the weights of the current Actor network and the weight of the current evaluation network , and then use the soft update method to update the weights of the two target networks respectively, namely: In the formula, It is a soft update parameter.

7. The driving and braking control method for a pure electric commercial vehicle driven by dual motors on front and rear axles according to claim 6, characterized in that: In step S6, a truncated triple Q learning strategy is adopted to obtain a more accurate Q value estimation; The objective function expression is as follows: In the formula, and Represents different network parameters and The Q output under is the next state, is a set of possible actions; mean represents the mean operation on the truncated Q value.

8. The driving and braking control method for a pure electric commercial vehicle driven by dual motors on front and rear axles according to claim 1, characterized in that: In step S7, the CyclicLR learning rate scheduler is introduced to adjust the learning rate for periodic decay; The reward function is defined as follows: And, define the boundaries in the learning process as follows: In the formula, and Indicates the upper and lower limits of the battery SOC, and Indicates the upper and lower limits of the battery health status. Indicates the upper limit of the first motor speed, Indicates the upper limit of the second motor speed, Indicates the upper limit of the torque of the first motor, Indicates the upper limit of the torque of the second motor.

9. The driving and braking control method for a pure electric commercial vehicle driven by dual motors on front and rear axles according to claim 1, characterized in that: In step S8, assuming that the speed prediction time domain is 3s, at the current time k, the vehicle speed is v, the historical speeds of k-3, k-2, k-1 and the current speed at time k are input into the speed prediction module, and the module outputs the predicted speeds at time k+1, k+2, k+3; then The vehicle speed information is input into the intelligent agent module of the learning predictive control for training, and the vehicle speed is used , battery SOC, SOH, and update the state of the dual-motor power system within the prediction time domain: Get the optimal action at time t Then, using the state quantity at time k and the best action , get the reward at time k , and the next state quantity , then the sequence Add to the sequence of experiences; The strategy network and value network of the improved DDPG algorithm are used for training to obtain the energy management model of a dual-motor pure electric vehicle based on learning model predictive control.

10. The driving and braking control method for a pure electric commercial vehicle driven by dual motors on front and rear axles according to claim 1, characterized in that: In step S10, after the training is completed, the energy management model of the learning predictive control is loaded into the vehicle controller, and the specific execution is as follows: During the actual driving process of the vehicle, the short-term historical speed of the vehicle and the current speed of the vehicle are input into the speed prediction model to solve the speed sequence of the vehicle in the prediction time domain in real time; Then, the vehicle speed in the predicted time domain is input into the trained DDPG agent. At each moment, the optimal control action, that is, the torque distribution coefficient of the front and rear motors, is obtained through calculation, and it is input into the motor controller to distribute the front and rear motor torques, realizing real-time rolling optimization.

Images